Despite advancements in technology, including deep learning techniques, Single-frame InfraRed Small Target (SIRST) detection in InfraRed (IR) imagery remains challenging, requiring further research and innovation. The lack of high-level semantic information causes small IR target features to diminish in the deeper layers of convolutional neural networks, reducing the network’s ability to accurately represent and identify these targets. This paper proposes a novel SIRST detection approach, STNet (Small Target detection Network), built on the U2-Net (Nested U-shape Network) architecture. STNet incorporates two key components: the MultiLayer Feature Fusion (MLFF) module and the Fast Fourier Block (FFB). The MLFF module enhances the model’s ability to integrate and leverage features from multiple layers, combining low-level details with high-level semantic information for more accurate SIRST detection. The FFB further improves the model’s performance by enabling feature extraction in the frequency domain, preserving small target features in the deep layers of the network, which enhances the performance of the detection process. Experimental results on the NUDT-SIRST and IRSTD-1K datasets show that STNet consistently outperforms other state-of-the-art methods. On the NUDT-SIRST dataset, STNet achieves the highest performance with an IoU of \(87.25\%\) , nIoU of \(87.23\%\) , and Pd of \(98.51\%\) , coupled with a low FA of \(5.92 \times 10^{-6}\) . Similarly, on the IRSTD-1K dataset, STNet achieves the best performance with \(72.04\%\) IoU, \(68.95\%\) nIoU, \(95.29\%\) Pd, and \(1.92 \times 10^{-6}\) FA. These results underscore STNet’s effectiveness in detecting and segmenting small IR targets in cluttered backgrounds.